Skip to content
Prev 9027 / 20628 Next
Original Raw

ZIP MCMCglmm

Katrin Ronnenberg 
Hi Jarrod, 
Thank you very much for your detailed help! I made some silly mistakes and had to think about some of your answers for a while. Unfortunately, these raise more questions. 
This is my most recent "working" r-code:

prior<-list(R=list(V=diag(2),nu=0.002,fix=2), G=list(G1=list(V=diag(2),nu=0.002)))

Mz<-MCMCglmm(sum_ppho ~ trait-1 + at.level(trait,1):year*
      at.level(trait,1):we+
      at.level(trait,1):(I(cdayofyear^2)+at.level(trait,1):dayofyear)+
      trait:sqkm,
random=~us(trait):EFF_Track, family="zipoisson",pr=TRUE,pl=TRUE,
      data=encounter, rcov=~idh(trait):units, prior=prior, verbose=FALSE,
      nitt=10000,thin =100, burnin = 1000)
summary(Mz)

This gives more or less the expected results, however, I'm not content with its implications.
First, I'm not sure if I misunderstand something, as you can see, I'm still fairly inexperienced. But probably my explanation of "sqkm" was not clear in my last email. 
It is a correction factor, which specifies for each transect "Eff_Track" it's effective observation area. For example in suboptimal weather conditions, you are only able to detect porpoises in shorter distances than in the next year in perfect weather conditions. So you decrease the observed area of transects in the correction factor. So I thought putting it in the random slope term, would account for that. Putting it as a fixed factor is probably of minor interest, as everyone would assume, that you see more porpoises in bigger survey areas. So basically it was my purpose to have the overall slope=0. This still leaves the problem that I need to allow for different slopes and intercepts for the zero-inflation and the poisson part. I tried fitting several terms but could not get the dimensions of the prior right. 

Also, my explanation of g(0) was a very simplified version of the real assumptions, I'm sorry for that. Yes, g(0) both corrects for unavailability bias (that could not be seen, because porpoises were diving) and the perception bias of those missed because of "sleeping" surveyors. Typically g(0) = <0.5, so more than half of the porpoises are missed by surveyors. Which means, quite some zeros will be due to false zeros. 
Does the ZIP model include any assumptions on the probability that animals are missed on transects with detections? I understood, it only accounts for the probability that a zero is a zero from the poisson process or a false zero. 
In fact we have an estimate of g(0) derived from distance sampling methods. I was wondering if it would be possible to use the known/estimated g(0) as starting value for a zero-inflation prior? So far I wasn't able to find any information on how to integrate starting values for latent variables. It seems like a fancy way of dealing with that problem, but maybe that will also exceed my skills. Or is there another way of dealing with that problem?
I believe correcting our counts directly will give even worse results. We give up the ?nteger properties of our variable and herewith change the distribution of data. Even worse with our original methods, we only corrected the counts, but not the zeros (by dividing through correction factors)
cheers,
Katrin


-------- Original-Nachricht --------